1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c16-neon-mlal.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <arm_neon.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c16__neon_mlal(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 int8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 16 * sizeof(int8_t));
44 int8_t* c0 = c;
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52 int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53 int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54 int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55 int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56 int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57 int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58 int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59 int32x4_t vacc1x0 = vacc0x0;
60 int32x4_t vacc1x1 = vacc0x1;
61 int32x4_t vacc1x2 = vacc0x2;
62 int32x4_t vacc1x3 = vacc0x3;
63 int32x4_t vacc1x4 = vacc0x4;
64 int32x4_t vacc1x5 = vacc0x5;
65 int32x4_t vacc1x6 = vacc0x6;
66 int32x4_t vacc1x7 = vacc0x7;
67
68 size_t p = ks;
69 do {
70 const int8_t* restrict a0 = a[0];
71 if XNN_UNPREDICTABLE(a0 != zero) {
72 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73 }
74 const int8_t* restrict a1 = a[1];
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77 }
78 a += 2;
79
80 // KC loop of 16 with up to 15 remainder
81 size_t k = kc;
82 while (k != 0) {
83 const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
84 const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
85
86 const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
87 const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
88 const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
89 const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
90 const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
91 const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
92 const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
93 const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
94
95 int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
96 int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
97 vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
98 vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
99 vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
100 vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
101 int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
102 int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
103 vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
104 vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
105 vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
106 vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
107 int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
108 int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
109 vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
110 vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
111 vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
112 vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
113 int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
114 int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
115 vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
116 vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
117 vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
118 vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
119 int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
120 int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
121 vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
122 vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
123 vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
124 vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
125 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
126 int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
127 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
128 vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
129 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
130 vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
131 int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
132 int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
133 vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
134 vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
135 vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
136 vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
137 int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
138 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
139 vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
140 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
141 vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
142 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
143
144 k -= 16 * sizeof(int8_t);
145 }
146
147 p -= 2 * sizeof(void*);
148 } while (p != 0);
149
150 #if XNN_ARCH_ARM64
151 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
152 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
153 const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
154 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
155 const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
156 const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
157 const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
158 const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
159 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
160 int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
161 int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
162 int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
163 #else
164 const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
165 const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
166 const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
167 const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
168 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
169 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
170 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
171 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
172 const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
173 const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
174 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
175 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
176 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
177 int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
178 const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
179 const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
180 const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
181 const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
182 const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
183 const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
184 int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
185 const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
186 const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
187 const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
188 const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
189 const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
190 const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
191 int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
192 #endif
193
194 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
195 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
196 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
197
198 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
199 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
200 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
201 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
202
203 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
204 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
205 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
206 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
207
208 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
209 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
210 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
211 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
212
213 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
214 #if XNN_ARCH_ARM64
215 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
216 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
217 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
218 #else
219 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
220 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
221
222 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
223 #endif
224 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
225 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
226
227 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
228
229 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
230
231 if (nc >= 8) {
232 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
233 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
234
235 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
236 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
237
238 a = (const int8_t**restrict) ((uintptr_t) a - ks);
239
240 nc -= 8;
241 } else {
242 if (nc & 4) {
243 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
244 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
245 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
246 }
247 if (nc & 2) {
248 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
249 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
250 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
251 }
252 if (nc & 1) {
253 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
254 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
255 }
256
257 nc = 0;
258 }
259 } while (nc != 0);
260 }
261